So how much VRAM do you actually need to run SDXL? OOM errors have scared off way too many people.

A lot of people load up SDXL in ComfyUI, slap on a ControlNet, hit generate, and ten seconds later their VRAM explodes. SDXL is way more resource-hungry than SD1.5—pick the wrong card and you’re either getting OOM every day or it’s so slow you wanna smash your keyboard. Here’s some real-world testing: a basic SDXL image at 1024 resolution uses about 6.

5G VRAM, each extra ControlNet adds 1.5 to 2.5G, upscalers pile on another 1 to 2G, and it stacks up fast. So if you want a smooth experience, 16G is the sweet spot—enough headroom for basic generation, single-layer ControlNet without choking VRAM, Tile upscaling or ESRGAN can handle it, and you can even train some LoRAs on the side.

At that price point, a few cards are within reach, and the 4070 Ti Super is one of the best bang-for-buck options. Don’t just look at whether it can run—SDXL workflows stack up, so leave enough buffer so you don’t run out of VRAM mid-generation.

Totally feel you on the ControlNet stacking eating up all the VRAM.

16G is the sweet spot, I agree with that.

6.5G is just the bare minimum, nowhere near enough for actual workflows.

Newbies always overlook leaving some headroom.

Totally agree on 16GB being the sweet spot. Even then, running two ControlNets still makes it tight—leaving some headroom really isn’t just empty talk.

1 Like